Clarify CRRC pre-compaction checkpoint need
This commit is contained in:
Meta Vens
parent
86ba6de5a9
commit
83a67ae2d0
4 changed files with 20 additions and 16 deletions
|
|
@ -113,6 +113,7 @@ These boundaries are more important than the individual method names:
|
|||
- `thread/epiphany/crrc` is a read-only coordinator recommendation over existing signals, not a scheduler, launch button, acceptance gate, compactor, or hidden state writer.
|
||||
- `thread/epiphany/coordinator` is a read-only fixed-lane MVP policy projection over existing signals, not a scheduler, launcher, acceptance gate, compactor, or hidden state writer.
|
||||
- Native CRRC automation may act only at safe turn-complete boundaries and only for coordinator-approved `compactRehydrateReorient` and `launchReorientWorker` actions. It must not auto-launch modeling or verification, auto-accept semantic findings, promote evidence, edit implementation code, or silently continue after unresolved drift.
|
||||
- Turn-complete-only CRRC automation is conservative but incomplete. The MVP still needs pre-compaction checkpoint intervention so the harness can interrupt or steer the active agent into banking scratch/checkpoint/map/evidence before compaction erases learned but unexternalized working context.
|
||||
- `thread/epiphany/roles` is a read-only role ownership projection, not a specialist scheduler, marketplace, launcher, acceptance gate, or second job backend.
|
||||
- `thread/epiphany/roleLaunch` is a bounded authority surface for fixed modeling/checkpoint and verification/review templates, not a broad scheduler or specialist marketplace.
|
||||
- `thread/epiphany/roleResult` is a read-only result projection, not a promotion gate, state writer, scheduler, or hidden continuation trigger.
|
||||
|
|
@ -213,8 +214,9 @@ proved the sequence-locked MVP policy across cold start, clean checkpoint,
|
|||
modeling, verification, CRRC drift/reorient, and high-pressure compact/dry-run
|
||||
paths. Native CRRC automation then wired the proved policy into turn-complete
|
||||
safe boundaries for compact and fixed reorient-worker launch only. The next MVP
|
||||
question is human operator testing, not another imagined organ: use the landed
|
||||
coordinator/status/artifact loop, then fix only concrete blockers.
|
||||
question is sharper: add the smallest pre-compaction checkpoint intervention so
|
||||
active work can be persisted before compaction, then dogfood the
|
||||
coordinator/status/artifact loop and fix only concrete blockers.
|
||||
|
||||
## Phase 6 Direction
|
||||
|
||||
|
|
@ -224,7 +226,7 @@ Useful candidates:
|
|||
|
||||
1. Put the fixed-lane coordinator MVP in front of a human operator through status/artifact review, then fix concrete usability blockers.
|
||||
2. Keep accepted worker findings review-gated; do not convert acceptance into automatic promotion of arbitrary worker output.
|
||||
3. Dogfood the landed safe-boundary CRRC automation on real work and fix only concrete blockers in compact/reorient-worker launch behavior.
|
||||
3. Add pre-compaction checkpoint intervention before dogfooding long work: detect rising pressure before the hard cliff, interrupt or steer into a bounded persistence pass, then compact/resume/reorient.
|
||||
|
||||
Do not spend Phase 6 polishing Phase 5 out of anxiety. The Phase 5 smoke harness
|
||||
is a regression guardrail, not a ritual drum circle for summoning more tiny
|
||||
|
|
|
|||
|
|
@ -32,6 +32,7 @@ Do not trust this file for the exact live HEAD. Always check git.
|
|||
- `tools/epiphany_mvp_status.py` is the first dogfood operator view. It starts or reads a thread through app-server and prints scene, pressure, reorient, jobs, roles, modeling/verification role result read-backs, reorient result, and CRRC recommendation as text or JSON.
|
||||
- `tools/epiphany_mvp_coordinator.py` is the first auditable fixed-lane coordinator runner. It starts or reads a thread through app-server, follows the harness-native coordinator action, can auto-launch modeling, verification, or reorient-worker jobs, keeps semantic findings review-gated by default, and writes summary, steps, rendered snapshots, transcript, stderr, and final next-action artifacts under `.epiphany-dogfood/coordinator` or a caller-provided artifact directory.
|
||||
- Native CRRC automation is now landed only at turn-complete safe boundaries. It may submit `Op::Compact` for coordinator-approved `compactRehydrateReorient`, and it may launch the fixed `reorient-worker` for coordinator-approved `launchReorientWorker`. It does not auto-launch modeling/verification, accept findings, promote evidence, edit implementation code, or keep going after reviewable semantic output.
|
||||
- This is not the full CRRC design. Turn-complete-only automation can still miss the important failure mode where the active agent has learned something during work but has not yet externalized it before compaction. The next CRRC slice should add a pre-compaction checkpoint intervention: detect rising pressure before the hard cliff, interrupt or steer the agent into a bounded persistence pass, then compact/resume/reorient only after scratch/checkpoint/map/evidence are banked as needed.
|
||||
- `tools/epiphany_mvp_dogfood.py` is the first auditable dogfood runner. It drives a bounded state/role/CRRC/reorientation loop and writes local artifacts under `.epiphany-dogfood/mvp-loop`, including raw app-server transcript, rendered snapshots, final status, vanilla-reference prompt/output, and comparison notes.
|
||||
- `tools/epiphany_mvp_live_specialist.py` is the first auditable live-specialist runner. It launches a real role specialist through `thread/epiphany/roleLaunch`, lets the spawned worker report through `report_agent_job_result`, reads it back through `thread/epiphany/roleResult`, and writes local artifacts under `.epiphany-dogfood/live-specialist`.
|
||||
- Internal `agent_job:` workers now get the reporting tool independent of the user-facing CSV spawn feature, and ephemeral worker sessions can initialize the sqlite state runtime on demand so specialist reports land in the shared backend.
|
||||
|
|
@ -289,17 +290,16 @@ The first live-specialist pass has also run. It produced
|
|||
inspected the smoke workspace, called `report_agent_job_result`, and
|
||||
`roleResult` projected a completed `checkpoint-ready` finding.
|
||||
|
||||
The fixed-lane coordinator MVP is now testable. Put it in front of a human
|
||||
operator through `tools/epiphany_mvp_coordinator.py`,
|
||||
`tools/epiphany_mvp_status.py`, `.epiphany-dogfood/coordinator`,
|
||||
`.epiphany-dogfood/mvp-loop`, and `.epiphany-dogfood/live-specialist`. Fix
|
||||
concrete blockers in that operator loop. Limited safe-boundary CRRC execution is
|
||||
already wired for compact and fixed reorient-worker launch actions, so the next
|
||||
unknown is how it behaves under real operator pressure, not which speculative
|
||||
organ to invent. CRRC is not a specialist-agent persona; the reorient-worker it
|
||||
may launch is the specialist. Do not turn the coordinator into a broad hidden
|
||||
dispatcher, arbitrary marketplace, alternate job backend, automatic semantic
|
||||
acceptance path, or GUI-as-source-of-truth.
|
||||
The fixed-lane coordinator MVP is testable, but the CRRC story is not complete.
|
||||
Limited safe-boundary CRRC execution is wired for compact and fixed
|
||||
reorient-worker launch actions, but it only acts after a turn ends. The next
|
||||
real organ is pre-compaction checkpoint intervention: when pressure rises before
|
||||
the hard cliff, interrupt or steer the active agent into a bounded persistence
|
||||
pass so learned working context is banked before compaction. After that, compact,
|
||||
resume, and reorient. CRRC is not a specialist-agent persona; the
|
||||
reorient-worker it may launch is the specialist. Do not turn the coordinator
|
||||
into a broad hidden dispatcher, arbitrary marketplace, alternate job backend,
|
||||
automatic semantic acceptance path, or GUI-as-source-of-truth.
|
||||
|
||||
Live `thread/epiphany/scene`, `thread/epiphany/jobs`, `thread/epiphany/roles`,
|
||||
`thread/epiphany/freshness`, `thread/epiphany/context`,
|
||||
|
|
|
|||
|
|
@ -61,3 +61,4 @@
|
|||
{"ts": "2026-04-28T16:56:56+00:00", "type": "design", "status": "ok", "note": "Clarified CRRC architecture: CRRC is harness workflow automation that watches context pressure and continuity, persists/checkpoints, triggers compact-rehydrate-reorient behavior, and may launch a bounded reorient-worker specialist; CRRC itself is not a specialist-agent persona."}
|
||||
{"ts": "2026-04-28T18:33:53+01:00", "type": "implementation", "status": "ok", "note": "Landed the fixed-lane Epiphany MVP coordinator: read-only thread/epiphany/coordinator composes roles, pressure, reorient, CRRC, role results, and reorient result into review-gated action recommendations, and tools/epiphany_mvp_coordinator.py drives the same policy through app-server APIs with auditable artifacts. Coordinator smoke covered cold prepareCheckpoint, clean continueImplementation, modeling review, verification review, source-drift reorient review, high-pressure compact/dry-run, and complete artifact bundles."}
|
||||
{"ts": "2026-04-28T18:56:47+01:00", "type": "implementation", "status": "ok", "note": "Landed limited native CRRC automation at turn-complete safe boundaries. It reuses the read-only coordinator policy and only auto-runs compactRehydrateReorient via Op::Compact or launchReorientWorker via the fixed reorient-worker job binding; modeling, verification, semantic acceptance, evidence promotion, implementation edits, and hidden continuation remain review-gated or manual. Focused coordinator automation assertions, job-progress event handling, coordinator smoke, status smoke, role smoke, and reorient-launch smoke passed."}
|
||||
{"ts": "2026-04-29T02:35:26+01:00", "type": "design", "status": "ok", "note": "Corrected the CRRC MVP target: turn-complete-only automation is insufficient because compaction can lose learned but unexternalized working context. The next CRRC slice should detect rising pressure before the hard cliff, interrupt or steer the active agent into a bounded persistence/checkpoint pass, then compact/resume/reorient after scratch/checkpoint/map/evidence are banked as needed."}
|
||||
|
|
|
|||
|
|
@ -86,6 +86,7 @@ accepted_design:
|
|||
- `thread/epiphany/coordinator` is the first harness-native fixed-lane coordinator surface: it composes roles, pressure, reorient, CRRC, roleResult, and reorientResult into a read-only action recommendation with target role, source signals, review requirements, automation eligibility, reason, and current lane summaries without mutating state or launching work.
|
||||
- `tools/epiphany_mvp_coordinator.py` is the first auditable coordinator runner: it drives the fixed single-user lanes through app-server APIs, may auto-launch modeling, verification, or reorient-worker jobs, keeps semantic findings review-gated by default, and writes coordinator summaries, JSONL steps, rendered snapshots, transcript, stderr, and final next-action artifacts.
|
||||
- Limited native CRRC automation is now wired only at turn-complete safe boundaries: it can submit a compact operation when the read-only coordinator recommends `compactRehydrateReorient`, and it can launch the fixed `reorient-worker` when the coordinator recommends `launchReorientWorker`. It does not auto-launch modeling or verification, auto-accept findings, promote evidence, edit implementation code, or continue after reviewable semantic results.
|
||||
- Turn-complete-only CRRC automation is not enough for the full Epiphany design. The next CRRC primitive must detect rising pressure before the hard compaction cliff and interrupt or steer the active agent into an explicit persistence/checkpoint pass so learned-but-unspoken working context is banked into scratch/checkpoint/map/evidence before compaction can sand it off.
|
||||
- `tools/epiphany_mvp_status.py` is the first dogfood operator view: it starts or reads a thread through app-server and renders scene, pressure, reorient, jobs, role result read-backs, reorient result, CRRC recommendation, and MVP role lanes as text or JSON without becoming a state store.
|
||||
- `thread/epiphany/roles` is the first harness-native role ownership surface: it projects implementation, modeling/checkpoint, verification/review, and reorientation lanes from authoritative typed state, bound jobs, pressure/reorient/CRRC, and worker result status without launching, scheduling, accepting, compacting, or mutating state.
|
||||
- `thread/epiphany/roleLaunch` and `thread/epiphany/roleResult` are the first explicit role-scoped specialist templates: modeling/checkpoint and verification/review can be launched through fixed bindings over the existing job-control seam and read back as reviewable findings, while implementation stays owned by the main coding agent and reorientation stays on the CRRC-specific launch/result path.
|
||||
|
|
@ -158,5 +159,5 @@ rejected_paths:
|
|||
reason: persistent state that accumulates every local truth becomes memory cruft and weakens rehydration
|
||||
|
||||
current_status:
|
||||
summary: Phase 1 durable state, Phase 2 prompt integration, Phase 3 typed read, Phase 4 retrieval/indexing, Phase 5 update/distill/propose/promote/read/notification, and Phase 6 read-only scene, jobs, roles, freshness, targeted context, pressure, reorient, CRRC, coordinator, reorient launch/result/accept, explicit modeling/checkpoint and verification/review role launch/result templates, MVP status, dogfood runner, coordinator runner, live specialist runner, and limited safe-boundary CRRC automation are landed. The coordinator endpoint and runner make fixed-lane modeling, verification, implementation continuation, and CRRC handoffs explicit through review-gated artifacts; native automation only compacts or launches the fixed reorient-worker at turn-complete boundaries when the coordinator says it can auto-run. Implementation remains main-agent work; CRRC remains workflow automation, not a specialist persona. Persistent state remains distilled; evidence is a belief ledger, not an activity feed.
|
||||
next_action: Dogfood the fixed-lane coordinator and native CRRC safe-boundary automation on real local work through `tools/epiphany_mvp_coordinator.py`, `tools/epiphany_mvp_status.py`, and their artifact bundles. Fix concrete operator-loop blockers only; do not build a broad ambient dispatcher, GUI-as-source-of-truth, arbitrary specialist marketplace, alternate job backend, or automatic semantic acceptance for the MVP.
|
||||
summary: Phase 1 durable state, Phase 2 prompt integration, Phase 3 typed read, Phase 4 retrieval/indexing, Phase 5 update/distill/propose/promote/read/notification, and Phase 6 read-only scene, jobs, roles, freshness, targeted context, pressure, reorient, CRRC, coordinator, reorient launch/result/accept, explicit modeling/checkpoint and verification/review role launch/result templates, MVP status, dogfood runner, coordinator runner, live specialist runner, and limited safe-boundary CRRC automation are landed. The coordinator endpoint and runner make fixed-lane modeling, verification, implementation continuation, and CRRC handoffs explicit through review-gated artifacts; native automation currently only compacts or launches the fixed reorient-worker at turn-complete boundaries when the coordinator says it can auto-run. That is conservative but incomplete: full CRRC needs a pre-compaction checkpoint intervention before active working memory is lost. Implementation remains main-agent work; CRRC remains workflow automation, not a specialist persona.
|
||||
next_action: Design and implement the smallest pre-compaction checkpoint intervention: when pressure is rising before the hard compaction cliff, interrupt or steer the active agent into a bounded persistence pass that writes scratch/checkpoint/map/evidence as needed, then compact/resume/reorient. Keep semantic acceptance review-gated, and do not build a broad ambient dispatcher, GUI-as-source-of-truth, arbitrary specialist marketplace, or alternate job backend for the MVP.
|
||||
|
|
|
|||
Reference in a new issue